openclaw - 💡(How to fix) Fix Allow depth-1 subagents to spawn skill-scoped sub-agents (remove sessions_spawn restriction at depth 1 [1 comments, 2 participants]

Summary

Enable sessions_spawn within subagent sessions (depth 1) so orchestrator agents can themselves delegate to skill-scoped sub-agents — supporting multi-level agent orchestration patterns.

Problem to solve

Currently, sessions_spawn is only available in the main agent session (depth 0). When a parent agent dispatches a specialist agent as a subagent (depth 1), that specialist cannot itself spawn further skill-scoped sub-agents — the tool is simply unavailable.

This breaks multi-level orchestration patterns where:

Main agent dispatches an orchestrator agent- That orchestrator is designed to fan out work across multiple skill-scoped executors in parallel Each executor loads a specific skill and returns a scoped result The orchestrator consolidates and returns one clean report to the main agent Today, depth-1 agents must collapse all their sub-tasks into a single session, defeating the purpose of skill-scoped isolation. The quality and context hygiene benefits of the pattern (isolated context per skill, parallel execution, clean consolidation) are lost.

Proposed behaviour: Allow sessions_spawn at depth 1, with depth-2 executors enforced as terminal (no further spawning). Depth limit: main → orchestrator → executor. Three levels maximum, hard-capped.

Proposed solution

Lift the sessions_spawn restriction at depth 1 with a hard depth cap of 2 (main → orchestrator → executor). Executors at depth 2 are terminal — sessions_spawn is unavailable to them, preventing unbounded recursion. The orchestrator at depth 1 receives the full tool set of the main session, including sessions_spawn, but executors it spawns do not. This requires no changes to the brief format or existing orchestration contracts — only the runtime availability of sessions_spawn needs to change per depth level.

Alternatives considered

No response

Impact

Context quality: Each executor runs with a clean, minimal context window scoped to one skill domain. No context bleed between domains — a secrets audit and a GDPR review do not share the same token budget or conversation history. Parallel execution: Independent skill domains can run simultaneously instead of sequentially in one bloated session. A 6-domain audit that currently takes one long sequential session becomes 6 parallel sessions converging at consolidation. Context rot mitigation: Long single-session audits accumulate noise — earlier findings pollute attention when reviewing later code. Isolated executor sessions start fresh per domain, eliminating this drift. Specialization fidelity: An executor briefed with one skill and one scope boundary produces higher-quality output than a generalist session context-switching across 10+ skill domains in sequence. Scalability: Orchestrators with many skills currently must either ignore skills or overload a single session. With depth-2 spawning, the number of skills an orchestrator can effectively apply scales with parallelism rather than context size.

Evidence/examples

An orchestrator agent with 11 security-related skills (OWASP, SAST, secrets management, GDPR, STRIDE, threat modelling, etc.) was dispatched as a depth-1 subagent. It attempted to fan out work across 4 parallel skill-scoped executors. sessions_spawn returned unavailable. The entire audit ran sequentially in one session — all 11 skills loaded, 19k output tokens, 6 minutes runtime. Parallel execution of 4 scoped sessions would have reduced runtime to ~2 minutes with cleaner per-domain output. The same limitation applies to performance auditors (Core Web Vitals + backend profiling + image optimization as separate executors), code reviewers (security pass + quality pass + spec compliance as separate executors), and any agent with 4+ skills whose work decomposes naturally into parallel scopes.

Additional information

No response